International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
Various reasons for a decreased phase correlation exist. Repeat 
pass interferometry is much more sensitive to temporal 
decorrelation, than optical stereoscopy to radiometric changes. 
Besides temporal decorrelation, thermal noise, SAR processing 
artefacts, baseline decorrelation, atmospheric artefacts and 
misregistration of the images have a decreasing impact on the 
correlation (Zebker et al., 1994). Layover causes total 
decorrelation of the signals, making the terrain also visible in 
the coherence map (Fig. 1, Fig. 2). In order to reduce statistical 
phase variations, the SAR multilook technique and 
interferogram filtering is applied. 
Fig. 1. DEM of test site 1. 
Fig. 2. Coherence map of test site 1. The 
coherence, respectively the height error, 
follows closely the terrain shape. 
Fig. 3. Correlation coefficient of SPOT stereo 
pair in test site 1. Correlation gaps occur 
randomly all over the site. 
To summarise, low cross-correlation indicates error suspicious 
parts in both DEMs and can be used as a common weight for 
the fusion process, as it leads to inaccurate measurements in 
both cases. 
In InSAR DEMs, height accuracy generally follows closely the 
phase coherence, but in transition areas from lower to higher 
coherence, a refinement of this global assumption is necessary. 
In case of LS phase unwrapping, isolated phase inconsistencies 
will be bridged, but highly correlated values adjacent to large 
areas showing low correlation may be error affected. This fact 
will be taken into account in the fusion procedure. 
Low correlation also leads to problems in stereo DEM 
generation, as it complicates the identification of conjugate 
points. Still, some mismatched points may have high correlation 
values, thus appearing as spikes in the DEM, and must be 
removed before the data fusion. 
4. DATA FUSION PROCESS 
As mentioned above, different types of errors occur in a DEM. 
Therefore, a DEM fusion by using cross-correlation would only 
reduce the statistic error, but leave systematic errors and 
blunders. In addition, unwanted holes, where correlation is low 
in both cases, would occur in the fused DEM. Therefore, the 
presence of a height estimate of the whole site is required, 
which will be refined by the fused InSAR and stereo-optical 
heights. A straightforward solution for this terrain estimate 
would be the filtered stereo DEM, which has a robust 
performance in most types of terrain and shows only small 
systematic errors. 
Median filtering is a common filter operation for the reduction 
of unavoidable outliers in stereo DEMs. Although it works well 
for a terrain smoothing, some problems remain: